search

brendanator / atari-rl

Atari - Deep Reinforcement Learning algorithms in TensorFlow
MIT License
135 stars 31 forks source link

Which is the fastest learner? #1

Closed ghost closed 6 years ago

ghost commented 7 years ago

Hi,

wow this is really a great repository :+1:

Which method do you recommend to try first?

The optimality tightening paper makes some really impressive claims, about the speed that they learn. I was just wondering if you've managed to reproduce any of their results? I just saw a theano implementation by one of the authors here

Thanks for your advice,

Aj

brendanator commented 7 years ago

Thanks Aj :)

Both Double-Q learning and Persistent Advantage Learning seem to work well so I would start with them along with the standard DQN algorithm

My first impression of optimality tightening is that it doesn't seem to give too much improvement, but I haven't tried it out on many games and it's quite possible that there's a bug somewhere. I haven't been trying to reproduce results so much as learn how all these algorithms work

Let me know if it works for you